Hyperspectral image feature extraction method based on global-local residual fusion network

Abstract

The present invention relates to the field of image processing technology, and more specifically, to a method for extracting hyperspectral image features based on a global-local residual fusion network. The method first extracts a primary feature map from a hyperspectral image. Global, local, and residual feature extraction are then performed on the primary feature map to obtain a global feature map, a local feature map, and a residual feature map. Finally, these three feature maps are fused to form a fused feature map. By employing three feature extraction methods, the method increases the diversity of the extracted features, thereby improving the quality of the resulting fused feature map.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a hyperspectral image feature extraction method based on a global-local residual fusion network. Background Art

[0002] Hyperspectral imagery (HSI) is a type of important image data composed of high-resolution spatial and spectral information. Hyperspectral images contain electromagnetic spectrum information ranging from visible light to near-infrared wavelengths and reflect the material information of objects in the image. The spectral and spatial features contained in hyperspectral images can be used to identify the category to which each pixel in the hyperspectral image belongs. This classification of pixels is called hyperspectral image classification. Hyperspectral image classification requires extracting feature maps from the hyperspectral image. Existing techniques for extracting feature maps involve relatively simple features, resulting in low quality, which makes the extracted feature maps unsuitable for hyperspectral image classification.

[0003] In summary, the feature maps extracted from hyperspectral images by existing technologies have low quality.

[0004] Therefore, the existing technology still needs to be improved and enhanced. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides a hyperspectral image feature extraction method based on a global-local residual fusion network, which solves the problem of low quality of feature maps extracted from hyperspectral images in the prior art.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a method for extracting hyperspectral image features based on a global-local residual fusion network, comprising:

[0008] Extracting primary feature maps of hyperspectral images;

[0009] Applying a global feature extraction algorithm to the primary feature map to obtain a global feature map;

[0010] Applying a local feature extraction algorithm to the primary feature map to obtain a local feature map;

[0011] Applying a residual network to the primary feature map to obtain a residual feature map;

[0012] The global feature map, the local feature map and the residual feature map are fused to obtain a fused feature map.

[0013] In one implementation, extracting a primary feature map of a hyperspectral image includes:

[0014] Applying a 1x1 convolution to the hyperspectral image to obtain a first primary feature map among the primary feature maps;

[0015] Applying a 3x3 convolution to the hyperspectral image to obtain a second primary feature map in the primary feature map;

[0016] A 5x5 convolution is applied to the hyperspectral image to obtain a third primary feature map in the primary feature maps.

[0017] In one implementation, applying a global feature extraction algorithm to the primary feature map to obtain a global feature map includes:

[0018] Applying a normalization algorithm to the first primary feature map to obtain a first normalized feature map;

[0019] Applying a normalization algorithm to the second primary feature map to obtain a second normalized feature map;

[0020] Applying a normalization algorithm to the third primary feature map to obtain a third normalized feature map;

[0021] Setting a first projection matrix for the first normalized feature map, a second projection matrix for the second normalized feature map, and a third projection matrix for the third normalized feature map;

[0022] Multiplying the first projection matrix by the first normalized feature map to obtain a first tensor;

[0023] Multiplying the second projection matrix by the second normalized feature map to obtain a second tensor;

[0024] Multiplying the third projection matrix by the third normalized feature map to obtain a third tensor;

[0025] Applying an attention mechanism to the first tensor, the second tensor, and the third tensor to perform global feature extraction to obtain a global feature map.

[0026] In one implementation, applying a local feature extraction algorithm to the primary feature map to obtain a local feature map includes:

[0027] splicing the first primary feature map, the second primary feature map, and the third primary feature map to obtain a spliced ​​primary feature map;

[0028] performing a convolution operation on the concatenated primary feature maps to expand the number of the concatenated primary feature maps to obtain a feature map set, wherein the feature map set is composed of each feature map group;

[0029] generating respective feature maps for each of the feature map groups;

[0030] multiplying each of the feature maps by a kernel weight to obtain a set of feature maps, wherein the kernel weight is related to a position of the feature map in the hyperspectral image;

[0031] Performing a shift transformation of a set operation kernel on each of the feature map sets to obtain each feature map after the shift transformation;

[0032] Performing an aggregation operation on each feature map after the shift transformation to obtain local feature information of each feature map group;

[0033] A splicing operation is performed on the local feature information of each feature map group to obtain a local feature map.

[0034] In one implementation, applying a residual network to the primary feature map to obtain a residual feature map includes:

[0035] splicing the first primary feature map, the second primary feature map, and the third primary feature map to obtain a spliced ​​primary feature map;

[0036] Compressing the concatenated primary feature map to a set size to obtain a compressed feature map, wherein the set size is equal to the size of the first primary feature map, the size of the second primary feature map, and the size of the third primary feature map;

[0037] A residual network is applied to the compressed feature map to obtain a residual feature map.

[0038] In one implementation, fusing the global feature map, the local feature map, and the residual feature map to obtain a fused feature map includes:

[0039] Adding the global feature map to the residual feature map to obtain a first residual map;

[0040] Subtract the local feature map and the residual feature map to obtain a second residual map;

[0041] Splicing the first residual map and the second residual map to obtain a spliced ​​feature map;

[0042] The spliced ​​feature maps are fused to obtain a fused feature map.

[0043] In one implementation, fusing the spliced ​​feature maps to obtain a fused feature map includes:

[0044] Performing normalization operations on the spliced ​​feature maps in sequence to obtain normalized spliced ​​feature maps;

[0045] Perform a 1x1 convolution operation on the normalized spliced ​​feature map to obtain the expanded spliced ​​feature map;

[0046] Perform a nonlinear operation on the expanded spliced ​​feature map to obtain a spliced ​​feature map after the nonlinear operation;

[0047] Compressing the spliced ​​feature map after the nonlinear operation to obtain a compressed spliced ​​feature map;

[0048] Performing a maximum pooling operation on the spliced ​​feature map to obtain a spliced ​​feature map after the maximum pooling operation;

[0049] The spliced ​​feature map after compression and the spliced ​​feature map after the maximum pooling operation are concatenated, and then a 1x1 convolution operation is performed to obtain a fused feature map.

[0050] Beneficial Effects: The present invention first extracts a primary feature map from a hyperspectral image. It then performs global feature extraction, local feature extraction, and residual feature extraction on the primary feature map to obtain a global feature map, a local feature map, and a residual feature map. Finally, these three feature maps are fused to obtain a fused feature map. Because the present invention uses three feature extraction methods, the diversity of the extracted features is increased, thereby improving the quality of the resulting fused feature map. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is the overall flow chart of the present invention;

[0052] Figure 2 This is a flow chart of extracting a splicing feature map in an embodiment of the present invention;

[0053] Figure 3 : is a structural diagram of the GLR-G in an embodiment of the present invention;

[0054] Figure 4 2 is a structural diagram of the GLR-L in an embodiment of the present invention;

[0055] Figure 5 : is a structural diagram of the GLR-R in an embodiment of the present invention;

[0056] Figure 6 This is a feature fusion flow chart in an embodiment of the present invention;

[0057] Figure 7 FIG. 4 is a structural diagram of an IBF in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The following is a clear and complete description of the technical solutions of the present invention in conjunction with the embodiments and the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0059] Research has found that hyperspectral imagery (HSI) is an important type of image data composed of high-resolution spatial and spectral information. Hyperspectral images contain electromagnetic spectrum information ranging from visible light to near-infrared wavelengths and reflect the material information of objects in the image. The spectral and spatial features contained in hyperspectral images can be used to identify the category to which each pixel in the hyperspectral image belongs. This classification of pixels is called hyperspectral image classification. The prerequisite for hyperspectral image classification is to extract the feature map of the hyperspectral image. The feature maps extracted by existing technologies involve relatively simple features, resulting in low quality of the feature maps, which makes the extracted feature maps unfavorable for hyperspectral image classification.

[0060] To address the above-mentioned technical problems, the present invention provides a hyperspectral image feature extraction method based on a global-local residual fusion network, which addresses the low quality of feature maps extracted from hyperspectral images in existing technologies. In specific implementation, a primary feature map of the hyperspectral image is first extracted; a global feature extraction algorithm, a local feature extraction algorithm, and a residual network are then applied to the primary feature map, respectively, to obtain a global feature map, a local feature map, and a residual feature map. Finally, the global feature map, the local feature map, and the residual feature map are fused to obtain a fused feature map. The present invention can improve the quality of the fused feature map.

[0061] Exemplary Methods

[0062] The hyperspectral image feature extraction method based on the global-local residual fusion network of this embodiment can be applied to a terminal device, which can be a terminal product with an image acquisition function, such as a computer. Figure 1 As shown in , the hyperspectral image feature extraction method based on the global-local residual fusion network specifically includes the following steps:

[0063] S100, extracting the primary feature map of the hyperspectral image.

[0064] like Figure 2 As shown, 1x1 convolution is applied to the hyperspectral image to obtain the first primary feature map I in the primary feature map ij *C (1) (The number is N, and the size of each first primary feature map is HxWxC / N, where H, W, and C represent the height, width, and number of channels of the hyperspectral image, respectively.ij is the hyperspectral image, C (1) is a 1x1 convolution kernel); 3x3 convolution is applied to the hyperspectral image to obtain the second primary feature map I in the primary feature map ij *C (3) (The number is N, and the size of each second primary feature map is HxWxC / N, C (3) is a 3x3 convolution kernel); 5x5 convolution is applied to the hyperspectral image to obtain the third primary feature map I in the primary feature map ij *C (5) (The number is N, and the size of each third primary feature map is HxWxC / N, C (5) is a 5x5 convolution kernel).

[0065] This embodiment uses three methods to extract primary feature maps, resulting in a diverse set of features. This also strengthens the inherent coupling between the different types of features extracted by different operators in the GLR module, effectively improving the accuracy of model classification (i.e., classifying pixels in the hyperspectral image based on the fused feature maps).

[0066] S200, applying a global feature extraction algorithm GLR-G to the primary feature map to obtain a global feature map G ij .

[0067] like Figure 3 As shown in the figure, GLR-G is mainly composed of an attention mechanism. GLR-G inputs three feature maps of size HxWxC / N and outputs a global feature map of size HxWxC / N.

[0068] In one embodiment, step S200 includes the following steps S201 to S208:

[0069] S201, applying a normalization algorithm to the first primary feature map to obtain a first normalized feature map bn(I ij *C (1) ,u( 1) ,σ (1) ,γ (1) ,β (1) ).

[0070]

[0071] Among them, u (1) , σ (1) , γ (1) , β (1) They are respectively the pixel value mean of the first primary feature map, the pixel value standard deviation of the first primary feature map, the normalized scaling factor, and the normalized deviation.

[0072] S202: Apply a normalization algorithm to the second primary feature map to obtain a second normalized feature map bn(I ij *C (3) ,u (3 ),σ( 3) ,γ (3) ,β (3) ).

[0073]

[0074] Among them, u (3) , σ (3) , γ (3) , β (3) are respectively the pixel value mean of the second primary feature map, the pixel value standard deviation of the second primary feature map, the normalized scaling factor, and the normalized deviation.

[0075] S203, applying a normalization algorithm to the third primary feature map to obtain a third normalized feature map bn(I ij *C (5) ,u (5) ,σ (5) ,γ (5) ,β (5) ).

[0076]

[0077] Among them, u (5) , σ (5) , γ (5) , β (5) They are respectively the mean pixel value of the third primary feature map, the standard deviation of the pixel value of the third primary feature map, the normalized scaling factor, and the normalized deviation.

[0078] S204: Setting a first projection matrix W for the first normalized feature map q , the second projection matrix W for the second normalized feature map k , the third projection matrix W for the third normalized feature map v .

[0079] W q 、W k 、W v Each matrix element in is a set constant.

[0080] S205: Multiply the first projection matrix by the first normalized feature map to obtain a first tensor q ij .

[0081] q ij =W q *bn(I ij*C (1) ,u (1) ,σ (1) ,γ (1) ,β (1) )

[0082] S206, multiplying the second projection matrix by the second normalized feature map to obtain a second tensor k ij .

[0083] k ij =W k *bn(I ij *C (3) ,u (3) ,σ (3) ,γ (3) ,β (3) )

[0084] S207, multiplying the third projection matrix by the third normalized feature map to obtain a third tensor v ij .

[0085] v ij =W k *bn(I ij *C (5) ,u( 5) ,σ (5) ,γ (5) ,β (5) )

[0086] S208, applying an attention mechanism to the first tensor, the second tensor, and the third tensor to extract global features, and obtaining a global feature map G ij .

[0087] The attention mechanism in this embodiment is a global feature extraction algorithm.

[0088]

[0089] In the formula, || represents the concatenation of the outputs of the N heads of the attention mechanism. patch_size In the (i,j) area, A(q ij ,k ab ) represents the use of Queries (i.e. ij ),Keys(i.e.k ij ) performs similarity calculation to obtain the attention update matrix, A(q ij ,k ab )*v ab Represents the attention update matrix as the weight update Values ​​(i.e. v ij ) process, G ij Cube patch_size(i,j,:) is the output obtained in the GLR-G module.

[0090] In this embodiment, the attention mechanism is used to extract feature maps, which has the following effects:

[0091] When extracting features, the attention mechanism tends to first focus on the important local information of objects, and then combine the information from different regions to form an overall impression of the observed object. The different levels of shallow feature information contained in the feature pool (consisting of the first primary feature map, the second primary feature map, and the third primary feature map) make the global feature extraction more targeted, thereby establishing a more reliable global dependency relationship.

[0092] S300: Apply a local feature extraction algorithm to the primary feature map to obtain a local feature map.

[0093] Local feature extraction algorithms such as Figure 4 As shown, the input of the local feature extraction algorithm is k 2 The input is a feature map of size HxWxC / N, and the output is a feature map of size HxWxC / N.

[0094] In the face of feature information of different receptive fields in the feature pool, the GLR-L submodule (local feature extraction algorithm) uses the convolution operator to extract the diversity of local information. In one embodiment, step S300 includes the following steps S301 to S307:

[0095] S301 : Concatenate the first primary feature map, the second primary feature map, and the third primary feature map to obtain a concatenated primary feature map.

[0096] Use Concat to concatenate the first primary feature map, the second primary feature map, and the third primary feature map, and the shape of the obtained concatenated primary feature map is (H×W×C / N)×3N.

[0097] S302: Perform a convolution operation on the concatenated primary feature maps to expand the number of the concatenated primary feature maps to obtain a feature map set, wherein the feature map set is composed of each feature map group.

[0098] Concatenate the primary feature maps and perform 1x1 convolution to expand them to a shape of (H×W×C / N)×k 2 ×N set of N feature maps l ij , where (H×W×C / N)×k 2 is a feature map group.

[0099] S303: Generate each feature map of each feature map group.

[0100] S304, multiply each of the feature maps by the kernel weight K p,q , get each feature map set The kernel weight is related to the position of the feature map in the hyperspectral image.

[0101] Generate k 2 feature maps to satisfy the GLR-L process with an operation kernel size of k. GLR-L will calculate the kernel weights K of different locations corresponding to specific locations (p, q) p,q For the input feature set l ij Perform corresponding linear projection to obtain the feature map set at each position where p,q∈{0,1,...,k-1}.

[0102]

[0103] S305, performing a shift transformation of the set operation kernel k on each of the feature map sets to obtain each feature map after the shift transformation

[0104]

[0105] in,

[0106] S306, performing an aggregation operation on each feature map after the shift transformation to obtain local feature information of each feature map group

[0107] GLR-L will get k 2 indivual Perform aggregation operations to obtain data blocks Cube patch_size A set of local feature information of (i,j,:)

[0108] S307, performing a splicing operation on the local feature information of each feature map group to obtain a local feature map L ij .

[0109]

[0110] S400: Apply a residual network to the primary feature map to obtain a residual feature map.

[0111] The process of residual network processing feature map is as follows Figure 5 As shown, the three feature maps, namely the first primary feature map, the second primary feature map, and the third primary feature map, are transformed into one feature map.

[0112] The residual network, the global feature extraction module, and the local feature extraction module constitute a residual structure that can solve problems such as gradient explosion and gradient loss.

[0113] The residual network (GLR-R) includes a feature map splicing module and a convolution module with a kernel size of 1. The feature map splicing module adopts the Concat splicing algorithm. The feature map splicing module splices the first primary feature map, the second primary feature map, and the third primary feature map to obtain a spliced ​​primary feature map with a shape of (H×W×C / N)×3N. The spliced ​​primary feature map is input into the convolution module to obtain a residual feature map with a shape of (H×W×C / N)×N.

[0114] In one embodiment, step S400 includes the following steps S401, S402, and S403:

[0115] S401 : Concatenate the first primary feature map, the second primary feature map, and the third primary feature map to obtain a concatenated primary feature map.

[0116] The so-called splicing is to put the above three primary feature maps together to form three feature maps.

[0117] S402 : Compress the concatenated primary feature map to a set size to obtain a compressed feature map, where the set size is equal to the size of the first primary feature map, the size of the second primary feature map, and the size of the third primary feature map.

[0118] The shape of the concatenated primary feature map is (H×W×C / N)×3N, and the shape of the compressed feature map is (H×W×C / N)×N.

[0119] S403: Apply a residual network to the compressed feature map to obtain a residual feature map.

[0120] The residual network directly outputs R ij , R ij That is (H×W×C / N)×N.

[0121] S500, fusing the global feature map, the local feature map and the residual feature map to obtain a fused feature map.

[0122] The feature map fusion process is as follows Figure 6 As shown in the figure, the whole fusion process mainly involves point convolution calculation and scaling calculation. Feature fusion is based on Figure 7 The nonlinear inverted bottleneck feature fusion module IBF shown in the figure performs feature fusion. The use of IBF has the following effects:

[0123] By moving the functional layer downward and retaining only nonlinear functions between dimensionality increase and dimensionality reduction operations, sufficient nonlinear changes can be made. Through sufficient nonlinear changes, fusion features with stronger feature semantic representation capabilities can be obtained, thereby more accurately fitting the mapping relationship between the extracted features and categories.

[0124] The IBF feature fusion module performs sufficient nonlinear changes and fusions on features at different levels, which can more accurately fit the mapping relationship between features and categories. In addition, IBF can also increase the depth of the entire network, allowing the model to express more abstract local or overall meanings of hyperspectral data.

[0125] In one embodiment, step S500 includes the following steps S501 to S509:

[0126] S501, add the global feature map to the residual feature map to obtain a first residual map G ij +R ij .

[0127] S502, the local feature map and the residual feature map are combined to obtain a second residual map L ij +R ij .

[0128] S503: Splice the first residual map and the second residual map to obtain a spliced ​​feature map.

[0129]

[0130] S504, performing normalization operations on the spliced ​​feature maps in sequence to obtain the normalized spliced ​​feature map M (out) .

[0131]

[0132] Where, for The variance of for , ε is a constant added for numerical stability, ω is a learnable scaling parameter, and δ is a learnable offset parameter.

[0133] S505: Perform a 1x1 convolution operation on the normalized spliced ​​feature map to obtain an expanded spliced ​​feature map.

[0134] Expansion is to quadruple the number of channels.

[0135] S506 , performing a nonlinear operation on the expanded spliced ​​feature map to obtain a spliced ​​feature map after the nonlinear operation.

[0136] The nonlinear operation is to deepen the nonlinearity of the feature through the nonlinear activation function RELU.

[0137] S507 , compressing the spliced ​​feature map after the nonlinear operation to obtain a compressed spliced ​​feature map.

[0138] The number of channels is compressed four times through convolution with a convolution kernel size of 1x1.

[0139] S508: Perform a maximum pooling operation on the spliced ​​feature map to obtain a spliced ​​feature map after the maximum pooling operation.

[0140] The maximum pooling operation mainly scales the values ​​on each channel through a learnable one-dimensional vector to automatically adjust the features and dynamically help the model improve the training quality.

[0141] S509: Concatenate the concatenated feature map after compression and the concatenated feature map after the maximum pooling operation, and then perform a 1x1 convolution operation to obtain a fused feature map.

[0142] The size of the fused feature map is (H×W×C / N)×N.

[0143] The fused feature map of this embodiment is a feature map obtained by combining the attention mechanism, convolution and residual structure. It fully mines the rich spatial-spectral feature information contained in the HSI data under the condition of limited training samples. The extracted features also have more advanced performance when facing classification tasks.

[0144] The following comparative experiments demonstrate that the fusion feature map extracted by the present invention is more conducive to the classification of pixels in the hyperspectral image:

[0145] As shown in Table 1, CCSU, YCSU, YRE, and YC are four standard HSI (hyperspectral image) data sets. SSLstm, Transformer, SCAT, ConvNext, U2Net, TwoCNN, RepVGG, Acmix, CTmixer, and the method of the present invention are used to extract fused feature maps on these four data sets. It can be seen from Table 1 that the pixel classification accuracy of the hyperspectral images extracted by the fused feature maps extracted by the method of the present invention reaches 93.43%, 90.87%, 92.14%, and 96.65% on the four data sets, respectively, which is much higher than the classification accuracy brought by the fused feature maps obtained by other methods.

[0146] Table 1

[0147]

Claims

1. A hyperspectral image feature extraction method based on a global-local residual fusion network, characterized in that: include: Extracting primary feature maps of hyperspectral images; Applying a global feature extraction algorithm to the primary feature map to obtain a global feature map; Applying a local feature extraction algorithm to the primary feature map to obtain a local feature map; Applying a residual network to the primary feature map to obtain a residual feature map; Fusing the global feature map, the local feature map, and the residual feature map to obtain a fused feature map; The step of extracting a primary feature map of a hyperspectral image comprises: Applying a 1x1 convolution to the hyperspectral image to obtain a first primary feature map among the primary feature maps; Applying a 3x3 convolution to the hyperspectral image to obtain a second primary feature map in the primary feature map; Applying a 5x5 convolution to the hyperspectral image to obtain a third primary feature map in the primary feature map; The applying a local feature extraction algorithm to the primary feature map to obtain a local feature map includes: splicing the first primary feature map, the second primary feature map, and the third primary feature map to obtain a spliced ​​primary feature map; performing a convolution operation on the concatenated primary feature maps to expand the number of the concatenated primary feature maps to obtain a feature map set, wherein the feature map set is composed of each feature map group; generating respective feature maps for each of the feature map groups; multiplying each of the feature maps by a kernel weight to obtain a set of feature maps, wherein the kernel weight is related to a position of the feature map in the hyperspectral image; Performing a shift transformation of a set operation kernel on each of the feature map sets to obtain each feature map after the shift transformation; Performing an aggregation operation on each feature map after the shift transformation to obtain local feature information of each feature map group; A splicing operation is performed on the local feature information of each feature map group to obtain a local feature map.

2. The hyperspectral image feature extraction method based on global-local residual fusion network according to claim 1, characterized in that: Applying a global feature extraction algorithm to the primary feature map to obtain a global feature map includes: Applying a normalization algorithm to the first primary feature map to obtain a first normalized feature map; Applying a normalization algorithm to the second primary feature map to obtain a second normalized feature map; Applying a normalization algorithm to the third primary feature map to obtain a third normalized feature map; Setting a first projection matrix for the first normalized feature map, a second projection matrix for the second normalized feature map, and a third projection matrix for the third normalized feature map; Multiplying the first projection matrix by the first normalized feature map to obtain a first tensor; Multiplying the second projection matrix by the second normalized feature map to obtain a second tensor; Multiplying the third projection matrix by the third normalized feature map to obtain a third tensor; Applying an attention mechanism to the first tensor, the second tensor, and the third tensor to perform global feature extraction to obtain a global feature map.

3. The hyperspectral image feature extraction method based on global-local residual fusion network according to claim 1, characterized in that: Applying a residual network to the primary feature map to obtain a residual feature map includes: splicing the first primary feature map, the second primary feature map, and the third primary feature map to obtain a spliced ​​primary feature map; Compressing the concatenated primary feature map to a set size to obtain a compressed feature map, wherein the set size is equal to the size of the first primary feature map, the size of the second primary feature map, and the size of the third primary feature map; A residual network is applied to the compressed feature map to obtain a residual feature map.

4. The hyperspectral image feature extraction method based on global-local residual fusion network according to claim 1, characterized in that: The fusing the global feature map, the local feature map, and the residual feature map to obtain a fused feature map includes: Adding the global feature map to the residual feature map to obtain a first residual map; Subtract the local feature map and the residual feature map to obtain a second residual map; Splicing the first residual map and the second residual map to obtain a spliced ​​feature map; The spliced ​​feature maps are fused to obtain a fused feature map.

5. The method for extracting hyperspectral image features based on a global-local residual fusion network according to claim 4, wherein: The step of fusing the spliced ​​feature maps to obtain a fused feature map includes: Performing normalization operations on the spliced ​​feature maps in sequence to obtain normalized spliced ​​feature maps; Perform a 1x1 convolution operation on the normalized spliced ​​feature map to obtain the expanded spliced ​​feature map; Perform a nonlinear operation on the expanded spliced ​​feature map to obtain a spliced ​​feature map after the nonlinear operation; Compressing the spliced ​​feature map after the nonlinear operation to obtain a compressed spliced ​​feature map; Performing a maximum pooling operation on the spliced ​​feature map to obtain a spliced ​​feature map after the maximum pooling operation; The spliced ​​feature map after compression and the spliced ​​feature map after the maximum pooling operation are concatenated, and then a 1x1 convolution operation is performed to obtain a fused feature map.

Images